AI/diffusers
1
0
Fork 0
mirror of https://github.com/huggingface/diffusers.git synced 2026-01-27 17:22:53 +03:00
Code Activity

[Diffusers -> Original SD conversion] fix things (#6933)

Browse Source 
* fix: bias loading bug

* fixes for SDXL

* apply changes to the conversion script to match single_file_utils.py

* do transpose to match the single file loading logic.
This commit is contained in:
Sayak Paul
2024-02-12 17:30:22 +05:30
committed by GitHub
parent 75aee39eac
commit 371f765908
3 changed files with 15 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
scripts/convert_diffusers_to_original_sdxl.py
View File

@@ -167,7 +167,10 @@ vae_conversion_map_attn = [
def reshape_weight_for_sd(w):
# convert HF linear weights to SD conv2d weights
return w.reshape(*w.shape, 1, 1)
if not w.ndim == 1:
return w.reshape(*w.shape, 1, 1)
else:
return w
def convert_vae_state_dict(vae_state_dict):
@@ -321,11 +324,18 @@ if __name__ == "__main__":
vae_state_dict = convert_vae_state_dict(vae_state_dict)
vae_state_dict = {"first_stage_model." + k: v for k, v in vae_state_dict.items()}
# Convert text encoder 1
text_enc_dict = convert_openai_text_enc_state_dict(text_enc_dict)
text_enc_dict = {"conditioner.embedders.0.transformer." + k: v for k, v in text_enc_dict.items()}
# Convert text encoder 2
text_enc_2_dict = convert_openclip_text_enc_state_dict(text_enc_2_dict)
text_enc_2_dict = {"conditioner.embedders.1.model." + k: v for k, v in text_enc_2_dict.items()}
# We call the `.T.contiguous()` to match what's done in
# https://github.com/huggingface/diffusers/blob/84905ca7287876b925b6bf8e9bb92fec21c78764/src/diffusers/loaders/single_file_utils.py#L1085
text_enc_2_dict["conditioner.embedders.1.model.text_projection"] = text_enc_2_dict.pop(
"conditioner.embedders.1.model.text_projection.weight"
).T.contiguous()
# Put together new checkpoint
state_dict = {**unet_state_dict, **vae_state_dict, **text_enc_dict, **text_enc_2_dict}

5
scripts/convert_diffusers_to_original_stable_diffusion.py
View File

@@ -170,7 +170,10 @@ vae_extra_conversion_map = [
def reshape_weight_for_sd(w):
# convert HF linear weights to SD conv2d weights
return w.reshape(*w.shape, 1, 1)
if not w.ndim == 1:
return w.reshape(*w.shape, 1, 1)
else:
return w
def convert_vae_state_dict(vae_state_dict):

1
src/diffusers/loaders/single_file_utils.py
View File

@@ -1112,7 +1112,6 @@ def create_text_encoder_from_open_clip_checkpoint(
text_model_dict[diffusers_key + ".q_proj.bias"] = weight_value[:text_proj_dim]
text_model_dict[diffusers_key + ".k_proj.bias"] = weight_value[text_proj_dim : text_proj_dim * 2]
text_model_dict[diffusers_key + ".v_proj.bias"] = weight_value[text_proj_dim * 2 :]
else:
text_model_dict[diffusers_key] = checkpoint[key]